SSD

SSD At the same time, we learn from YOLO The thought and of network Faster R-CNN Of anchor Mechanism , send SSD It can not only make rapid prediction, but also obtain the position of the target relatively accurately .

(b) and (c) They represent different feature layers , chart c Relative to the graph b Close to the final prediction result , Therefore, it can detect more targets across the same number of pixels . Such as b Shown , At each node of the feature layer, there will be 4 A different size anchor(1:1 Two ,1：2 Two ）, Such as c Shown , The same is true at the feature level . According to the true target rectangle with each anchor Of IOU Size calculation shows that ,b Contained in the 2 individual anchor Positive sample ,c There are only 1 individual . 

SSD Some characteristics of ：

1. Multi scale feature layer is used for detection . stay Faster R-CNN Of RPN in ,anchor It is generated on the last feature layer of the dry network , And in the SSD in ,anchor Not just on the last feature layer , And at several high-level feature layers, there are also anchor. The size of these layers decreases in turn , bring SSD It can detect targets with different scales .
2. SSD Generated by all feature layers in anchor Will be screened by positive and negative samples , Then classify the scores and Bbox Position learning . That is, the positive and negative samples generated on the feature layer will be directly classified （ClassNum Categories ） as well as Bbox Learning from , Unlike Faster R-CNN In that way, first learn whether there are objects in the first step （ Only 0/1 Two categories ） as well as Bbox Location , Then learn the final classification in the second step （ClassNum Categories ） And right Bbox Fine adjustment of position .

In practice , We should not only focus on accuracy , In many cases, speed is also considered , For example, real-time detection of video content , At this time, we hope that the method can balance the speed and accuracy well .